Home / Benchmarks

Benchmarks

Public results for the parts we build with. They are not EMD Lab measurements; our own lab results will be added as testing is completed.

Workstation graphics card ready for testing

Expected performance

Two graphics cards with similar specs can perform very differently in the same application. These ranges show how the parts we build with compare in real software.

Coming from the EMD Lab

DaVinci ResolveExport and effects
BlenderGPU and CPU rendering
Local LLM inferenceTokens per second, memory use

Each lab result will include:

Exact CPU, GPU and memory
Software and driver versions
Test date and method
Several runs, not one

AI deep dive

What more GPU memory gets you

The model has to fit in GPU memory, with room left for the conversation. More memory means a bigger model, a longer document, or more people at once. Figures below are for 4-bit models (Q4), a 32K-token conversation and one user. The “Local AI” tab above shows measured community speeds for a 7B model; these are estimates for larger models.

1 x GeForce RTX 509032 GB GPU memory, about 31.4 GB usable

Largest model that fits: SEA-LION v4.8 30B-A3B (Nemotron)27.6 GB needed, about 200 tokens/s
Largest dense model that fits: Gemma 4 31B24.9 GB needed, about 55 tokens/s
Chats at once (Qwen3.8 27B, 32K): 5fit in memory

2 x GeForce RTX 509064 GB GPU memory, about 62.8 GB usable

Largest model that fits: Qwen3-Coder-Next (80B-A3B)52.7 GB needed, about 233 tokens/s
Largest dense model that fits: Llama 3.3 70B Instruct56.3 GB needed, about 37 tokens/s
Chats at once (Qwen3.8 27B, 32K): 18fit in memory

1 x RTX PRO 500072 GB GPU memory, about 71.4 GB usable

Largest model that fits: Llama 4 Scout 17B-16E (109B-A17B)71.3 GB needed, about 74 tokens/s
Largest dense model that fits: Llama 3.3 70B Instruct55.7 GB needed, about 20 tokens/s
Chats at once (Qwen3.8 27B, 32K): 22fit in memory

1 x RTX PRO 600096 GB GPU memory, about 95.4 GB usable

Largest model that fits: NVIDIA Nemotron 3 Super 120B-A12B90.3 GB needed, about 112 tokens/s
Largest dense model that fits: Llama 3.3 70B Instruct55.7 GB needed, about 26 tokens/s
Chats at once (Qwen3.8 27B, 32K): 33fit in memory

2 x RTX PRO 6000192 GB GPU memory, about 191 GB usable

Largest model that fits: Inkling-Small (276B-A12B)170 GB needed, about 149 tokens/s
Largest dense model that fits: Llama 3.3 70B Instruct56.3 GB needed, about 26 tokens/s
Chats at once (Qwen3.8 27B, 32K): 74fit in memory

4 x RTX PRO 6000 (Team AI server)384 GB GPU memory, about 382 GB usable

Largest model that fits: Qwen3.5-397B-A17B256 GB needed, about 153 tokens/s
Largest dense model that fits: Llama 3.3 70B Instruct57.5 GB needed, about 26 tokens/s
Chats at once (Qwen3.8 27B, 32K): 156fit in memory

Mixture-of-experts models (explained below) use only part of the model for each word, so the largest model is often much faster than the largest dense model. Models with a non-commercial licence are left out. “Chats at once” counts how many 32K conversations with Qwen3.8 27B fit in memory together. That is memory, not speed: with many people at once, each gets fewer tokens per second.

Estimates from EMD Lab's AI selector (memory from model files; speed from memory bandwidth, calibrated on 25 public benchmarks). Not measurements.

What a bigger model buys you

Open models come in size classes. A bigger model needs more memory, and a bigger dense model also runs slower. Here is what each class asks for, using models from our AI selector.

Small

Up to about 20B parameters. Quick answers, drafting, summaries, simple assistants.

Qwen3.5 9B9.7 B

Best small all-rounder, with official Bahasa Indonesia support.

Needs about 8.7 GB (Q4, 32K). Smallest machine: 1 x GeForce RTX 5070 Ti (16 GB), about 79 tokens/s.

gpt-oss-20b20.9 B total, 3.6 B active

OpenAI's open model: fast, good with tools, fits a 16 GB GPU.

Needs about 14 GB (Q4, 32K). Smallest machine: 1 x GeForce RTX 5070 Ti (16 GB), about 147 tokens/s.

Sahabat-AI v1 9B (Gemma 2)9.2 B

Indonesian-first model from GoTo & Indosat. Short 8K context.

Needs about 13.5 GB (Q4, 32K). Smallest machine: 1 x GeForce RTX 5070 Ti (16 GB), about 53 tokens/s.

Medium

About 24B to 36B. The usual sweet spot for one person: coding, documents, reasoning.

Qwen3.8 27B27.8 B

Latest Qwen 27B: top dense model for a single 32 GB GPU, official Bahasa Indonesia.

Needs about 21.5 GB (Q4, 32K). Smallest machine: 1 x GeForce RTX 5090 (32 GB), about 63 tokens/s.

Qwen3.6 35B-A3B (MoE)36 B total, 3 B active

Fast MoE all-rounder that uses very little memory for long documents.

Needs about 24.9 GB (Q4, 32K). Smallest machine: 1 x GeForce RTX 5090 (32 GB), about 210 tokens/s.

Qwen3-Coder 30B-A3B Instruct30.5 B total, 3.3 B active

The most-downloaded local coding model.

Needs about 23.5 GB (Q4, 32K). Smallest machine: 1 x GeForce RTX 5090 (32 GB), about 153 tokens/s.

Large

About 70B to 180B. Harder reasoning, long documents, agent and coding work.

gpt-oss-120b117 B total, 5.1 B active

OpenAI's flagship open model: strong reasoning, fast for its size.

Needs about 67.1 GB (Q4, 32K). Smallest machine: 1 x RTX PRO 5000 (72 GB), about 152 tokens/s.

Llama 3.3 70B Instruct70.6 B

Classic 70B dense model: capable, but heavy and slow for its quality today.

Needs about 56.3 GB (Q4, 32K). Smallest machine: 2 x GeForce RTX 5090 (64 GB), about 37 tokens/s.

Qwen3-Coder-Next (80B-A3B)80 B total, 3 B active

Best local coding model in 2026, fast for its size.

Needs about 52.7 GB (Q4, 32K). Smallest machine: 2 x GeForce RTX 5090 (64 GB), about 233 tokens/s.

Extra large

About 230B to 400B. Near the best open models. Needs a multi-GPU machine.

Qwen3.5-397B-A17B397 B total, 17 B active

Qwen flagship class with good Bahasa Indonesia. Needs 256 GB or more.

Needs about 256 GB (Q4, 32K). Smallest machine: 4 x RTX PRO 6000 (384 GB), about 153 tokens/s.

DeepSeek-V4-Flash-0731 (284B-A13B)284 B total, 13 B active

DeepSeek's efficient big model, near-frontier quality on a 256 GB machine.

Needs about 162 GB (Q4, 32K). Smallest machine: 2 x RTX PRO 6000 (192 GB), about 156 tokens/s.

Bigger usually means more capableLarger models usually follow complex instructions better, write better code and handle Indonesian better. They also need more memory and answer more slowly.

Dense or mixture-of-expertsA dense model uses all of its parameters for every word. A mixture-of-experts (MoE) model has many “experts” and uses only some of them for each word, so it runs faster than its size suggests. It still has to fit in memory in full.

Active and total parametersFor MoE models we show both: total parameters set the memory needed, active parameters set the speed.

Estimates from EMD Lab's AI selector (memory from model files; speed from memory bandwidth, calibrated on 25 public benchmarks). Not measurements. Speeds are for one user, a 4-bit model and a 32K conversation. See all 36 models, with a memory calculator, on the AI models page.

RTX PRO or GeForce RTX 5090? A different stage, not a better card.

The GeForce RTX 5090 is an excellent card for AI, and for one person exploring models it is often the right choice. RTX PRO cards are built for what comes next: bigger models, longer runs, shared use. Pick by the stage you are in.

Stage 1

Explore

One person trying models up to about 30B: coding help, document questions, an assistant of your own. One GeForce RTX 5090 (32 GB) is the fast, sensible choice.

Configure an RTX 5090 workstation

Stage 2

Produce

Bigger 70B to 120B models, long documents, long unattended runs. This is where RTX PRO fits:

48 to 96 GB per card, with ECC memory
NVIDIA RTX Enterprise drivers and software
300 W Max-Q cards that fit two to four in one machine

Configure an AI workstation Find the right size

Stage 3

Serve a team

Several people using the same model at once. The Team AI server has four RTX PRO 6000 Max-Q cards and 384 GB of GPU memory.

Every RTX PRO 6000, the Max-Q included, can use Multi-Instance GPU (MIG) to split one card into up to 4 fully isolated GPUs, each with its own memory.

See the Team AI server Configure it

Speed is not the difference. Both cards have 1,792 GB/s of memory bandwidth, so when a model fits in 32 GB their speeds are close. The measured 7B results on the “Local AI” tab agree: RTX 5090 290 to 305 tokens/s, RTX PRO 6000 Blackwell 270 to 285 tokens/s. The difference is memory size: 32 GB against 96 GB per card.

GeForce RTX 5090

32 GB GDDR7, 512-bit bus
575 W total graphics power; NVIDIA lists 1,000 W required system power
Founders Edition is 2-slot and 304 mm long; partner cards in our catalog are 3.5 slots and 360 mm
3 x 9th-gen NVENC, 2 x 6th-gen NVDEC
Game Ready and Studio drivers
No NVLink

RTX PRO 6000 Blackwell Workstation Edition

96 GB GDDR7 with ECC, 1,792 GB/s
600 W, dual slot, double flow-through cooler
4 x NVENC, 4 x NVDEC, 4 x DisplayPort 2.1
Multi-Instance GPU (MIG): one card as up to 4 fully isolated GPUs
NVIDIA RTX Enterprise drivers and software

RTX PRO 6000 Blackwell Max-Q

The same 96 GB with ECC and 1,792 GB/s
300 W, dual slot, blower-style (active) cooling
Multi-Instance GPU (MIG): one card as up to 4 fully isolated GPUs
NVIDIA says it scales from one to four GPUs in one workstation

Card specifications: nvidia.com product pages, October 2026. SolidWorks and Creo officially support certified RTX PRO cards; GeForce works without that certification. Not sure which stage you are in? Try the AI selector or ask us.